Automated Text Coherence Analysis through Semantic Subspace Modelling and Ensemble Learning

Coherence is fundamental to producing clear and logical academic writing. Despite its importance, most automated evaluation systems primarily target grammar and style, often overlooking this critical discourse-level quality. This study presents a comprehensive comparative analysis of multiple computational approaches like classical machine learning, recurrent and convolutional neural networks (RNNs and CNNs), transformer-based architectures (BERT, RoBERTa, SBERT), and graph-based and hybrid models, specifically for coherence assessment in academic texts. Using a labelled academic writing dataset, model performance was benchmarked using accuracy, precision, recall and F1-scores to evaluate both training and test generalisation. Among classical models trained on TF-IDF and topic-based features, ensemble methods, such as Random Forest and XGBoost, demonstrated stronger robustness than linear and probabilistic baselines, while the nonlinear Multi-Layer Perceptron (MLP) achieved a test F1-score of 76.97%. Deep neural models exhibited stable performance, with the BiLSTM achieving a 76% test F1-score, outperforming LSTM, GRU and CNN models. Graph-based approaches integrating knowledge graph features improved generalisation, reaching a 76.64% test F1-score. The best overall performance was achieved by a hybrid model combining RoBERTa embeddings with Random Forest, attaining an 81.64% test F1-score. These findings indicate that integrating deep semantic embeddings with well-regularised ensemble classifiers offers the most effective and interpretable strategy for automated coherence evaluation.